What we've learned shipping production systems. AI, architecture, and 25 years of pattern recognition.
Before spending on fine-tuning, we measured whether a 1.7k-word AI constitution actually produces different behavior than a 29k-word one. The cheap probe answered it, and located the bottleneck.
A wind and solar forecaster we built on the side, mostly by pointing AI agents at the problem, is now live in production on Predico, the platform used by Belgium's grid operator, and gets scored against professional forecasters. Here is the honest scorecard, losses included.
We stress-tested whether a compact 6-principle AI constitution (Kardia) and Anthropic's ~29,000-word hierarchy actually pick different answers — using two independent strong judges. The divergence is real; the story about where it lives is judge-dependent. An honest interim report.
Before fine-tuning anything, we measured whether Kardia's short Citadel constitution and Anthropic's long hierarchy constitution encode different preferences on the same response pairs. Weak judges hid the signal; strong cross-family judges found a thin, real gap — about 17% opposite picks on a hard probe set. Method, confounds, and what we are not claiming.
We gave Claude Code a single /goal prompt. It recovered a 1997 undergraduate thesis from the Internet Archive, ported its solver, built a classic Mastermind board, and published it — with its thought process included.
Augur forecasts wind and solar for three European countries and scores every forecast against the grid operator's own — including where it loses. Why the baseline you beat is the whole story, and what honest energy forecasting actually looks like.
What happens when an AI agent reads a scenario document and executes it on a real iOS simulator — adapting to layout shifts, verifying with screenshots, and writing the report itself.
Your team ships 5x the code. Who's checking it? AI-generated code has specific failure modes that junior engineers can't catch. Here's the honest picture.
After building agentic AI systems in production environments, here's the honest accounting of what works, what fails, and what the hype gets wrong. From Eloquentix engineers.
Model Context Protocol connects your AI to live business data. Before you deploy: prompt injection, tool poisoning, over-permissioning, and a four-pillar security framework from Eloquentix engineers.
12 years building E.ON's Virtual Power Plant — dispatching 400 assets simultaneously across four countries. What the architecture looks like, what failed, and what it teaches about real-time systems at TSO grade.
While talking to someone who is among the most forward implementors of AI solutions in the enterprise, we both agreed that where agents are deployed is...
We just published Kardia, our take on what an AI constitution should be. kardia.eloquentix.com
In mid-February, after coming back from China and catching up with AI, a partner told me about all the agents he had set up, the scaffolding, the...
The divergence between those who are still on GPT's creepy voice and those running agents of different sorts is widening, and it may be that it is not...
May I kindly ask to share a new term: RandomMaxxing.
Some of Claude Mythos' adventures (credit @hesamation). This is exactly why a constitution built on formation, not just constraint, matters:
Many who are aware of AI capabilities today are scared of their future. What is resilient in an AI world? Skills? Judgement? Rolodex? Knowledge? All...
You may have heard various tech voices say that we are headed to an age of abundance. This idea may seem cuckoo, but it is worth understanding where it...
I saw a post on X about an interaction between Jony Ive and Steve Jobs that stuck with me. Asking the models what they think landed on remarkable...
I was listening to the original iPod presentation from 2001, the '1000 songs in your pocket' one. Actually the most important part was at the beginning.
Our brain's capacity to forget is adapting to feeds. Just like you don't remember the details of the asphalt and trees that you saw in the last 24...
Most people have never heard of ZF. This German company quietly builds the world's best automatic transmissions. Their 8-speed ZF 8HP powers BMWs,...
After a few days taking to a young researcher at Wisconsin, he said that our discussions highlighted to him the difference between ML and what we now...
Singularity stuff. Karpathy just addressed the priciest bottleneck in ML research for free. Researcher iteration speed. A senior ML engineer bleeds...
Our latest open-source contribution at Eloquentix: the Metals LSP plugin for Claude Code. This integration brings advanced Scala language server...
If it was not obvious.
Here is what Karpathy is saying and it hard to disagree after using tools in February 26. And these again are the worse tools we will ever have.
Google just released the Universal Commerce Protocol -- an open-source standard for AI-powered shopping.
We asked Claude what it thought about human code and how it would approach coding if humans were not in the loop. Here is a part of its answer.
From an Eloquentix developer:
Co-workers are non-deterministic. That is why we have audits.
There's a pattern in how nations lose economic dominance that we keep missing. Europe gave us quantum mechanics, radioactivity, and modern chemistry....
About Claude Code: The best compliment one can give to it is that it feels like playing a video game.
AI has done something no human ever could.
All the big labs pretrain on the ~same internet, but each has a secret sauce:
Congrats PER.
"The UX for long running AI Agents is going to be one of the most interesting design questions in the coming years. The more the agent is doing complex...
Having worked on the best VPP in Europe for the last 10 years, what Spain's massive blackout, exposed the risks of relying heavily on renewables...
GPT-4.1 is out and here are some tips from inside OpenAI.
Emerging UX principles for agentic AI -- systems that act autonomously to achieve goals -- focus on balancing user control, trust, and seamless...
With full appreciation of the clarity of his prose, here is Tobi Lutke's message to Shopify.
"radu, nobody understands your posts, so say whatever" here you go:
Garry Tan - yCombinator:
Grok3 Think mode is impressive. Since ChatGPT came out 27 months ago, at every model release I have tested it with my own test, a roman game that 1) Is...
These Cursor mfs forked VSCode, added anthropic keys & a chat window, and now are worth ~$700mm EACH at 25 years old.
Listening to AAPL earning, I heard that their services revenue is running at 105b/year with a 14% growth rate. Double clicking on that, they have over...
AI Agents & Microservices: A New Paradigm for Dev Teams
With GPT-4b micro, OpenAI shows that transformers are perfectly suited for protein engineering. Think about it - transformers process sequences of...
Summary of how DeepSeek V3 was so efficient at training a frontier-level model according to Perplexity:
Minimizing layers in an organization is empowering. It treats others like adults. This is rare especially outside the United States.